Anthropic explores how language models organize reasoning across layers

Researchers at Anthropic published findings on how Claude's internal representations function like a shared workspace, where different computational layers exchange information. The discovery could help developers and researchers better understand model behavior and potentially improve how AI systems approach complex reasoning tasks.